Author: Bruce Moreland
Date: 10:05:56 06/12/98
Go up one level in this thread
On June 12, 1998 at 01:21:27, Roberto Waldteufel wrote: > >Do you think the PV is actually worth the overhead of gathering it, >bearing in mind that the information in the hash table, ie the killer >moves stored with each node, effectively allow the PV from the previous >iteration to be quickly retrieved in the current search? When I first >implemented PV in a standard alpha-beta search without hashing (it was >on a machine with hardly any RAM), I found a speed-up of about 40% by >following this line first. Later, when I had more RAM to play with, I >implemented a hash table, which speeded up the search a bit more. >However, with the hash table working, I found that the search didn't >need the PV list any more, because it almost always found it from the >table, and so now I only bother to return best move at the root node and >I find the search increases in speed by another 5-10% or so. Of course >there is the risk that a PV node may have been overwritten, but this >occurs too rarely to hurt performance as much as maintaining the PV >separately did, particularly if the hash table is large. Does anybody >else use the hash table rather than the PV in this way? I know this is how some programs work, but I haven't tried it, since I have been gathering the PV from day one. I'd be concerned that I'd overwrite the table in cases where I have a non-huge table and a non-small amount of time. It would seem like I'd over-write the PV pretty consistently unless I was very careful. It doesn't take 5-10% overhead to record the PV, and I wonder how you could have arrived at this figure. bruce
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.